Pro-active agents with recurrent neural controllers

Embodied cognitive science is regarded as a bottom-up approach within artificial intelligence. It envisages the realisation and study of increasingly complex artificial agents. As a consequence, the research emphasised reactive agents, i.e., agents that always respond in the same way to the same sensory inputs. To reach higher agent capabilities, the research focus currently shifts from reactive to pro-active agents. The actions of pro-active agents do not only depend on the inputs, but also on the “internal state”. The internal state of an agent is defined as the subset of variables that co-determine the future input-output mapping of the agent. Typically, a pro-active agent has a neural controller. In this thesis, we investigate three mechanisms that realise an internal state in a neural controller: recurrency, neural inertia, and varying time delays. Since the internal state determines the agent’s behaviour in the environment, it influences the agent capabilities. At the moment it is a challenging question how the different mechanisms that realise an internal state determine the capabilities of pro-active agents. This leads us to the following problem statement. How do the mechanisms that realise an internal state influence the agent capabilities? To structure our investigations, we formulate two research questions: 1. What are the capabilities of agents with different types of recurrent neural controllers? 2. How are these capabilities related to the mechanisms realising an internal state? To answer the research questions, the following five types of recurrent neural controllers are investigated: the standard recurrent neural network, the recurrent neural network with a delta, the continuous-time recurrent neural network, the nonlinear autoregressive model with exogeneous inputs, and the time delay recurrent neural network. We employ the different types of recurrent neural networks as controllers of agents that have to perform robotic tasks. Following the methodology of evolutionary robotics, we performed experiments on the following three tasks: (1) the nest-finding task, (2) the driving task, and (3) the self-localisation task. All three tasks require an internal state to be solved.